Internalizing Agent Experience into Diffusion Model Weights via On-Policy Context Distillation

# Internalizing Agent Experience into Diffusion Model Weights via On-Policy Context Distillation
Source: [https://arxiv.org/html/2610.07250](https://arxiv.org/html/2610.07250)
Wenxuan Wang Zekai Liu Weinan Zhang Yu Cheng Yang YangHarbin Institute of Technology Shanghai AI Laboratory Shandong UniversityNanyang Technological University Shanghai Jiao Tong University\*Equal contribution\.†Corresponding author\.

###### Abstract

Wrapping an image generation model in an agentic harness can effectively boost Text\-to\-Image task performance: the harness can leverage memory, skills, workflow orchestration, result verification, and iterative refinement to continually construct and revise prompts, thereby eliciting better images\. These gains, however, remain external to the diffusion model and are realized only while the full harness runs\. We propose Diffusion On\-Policy Context Distillation \(D\-OPCD\), which treats the agent\-improved prompt as privileged context and distills the knowledge encoded in the agent harness into the weights of the diffusion model, so that the model retains part of the harness’s benefit when conditioned on the original query alone\. Using a Text\-to\-Image agent equipped with our proposed Auto Skill Evolver \(ASE\), we show that D\-OPCD can internalize harness capabilities into the generator’s weights, raising the average direct\-generation score from 60\.52 to 65\.09 across four benchmarks\. With this knowledge absorbed into the weights, the harness can shed its saturated skills and resume evolving: a second ASE round on the updated generator improves on a skill\-free harness by additional 1\.83 points, pointing toward text\-to\-image systems in which harness and model keep improving each other through continual co\-evolution\.[Code](https://github.com/Yummytanmo/D-OPCD-CoEvolution)is publicly available\.

## 1Introduction

Agentic harnesses have become an effective way to improve text\-to\-image \(T2I\) generation\. A harness wraps a fixed generator with reasoning, memory, skills, and verification: within a task, it reasons about and iteratively refines its generations\([Wang et al\., 2025](https://arxiv.org/html/2610.07250#bib.bib2);[He et al\., 2026](https://arxiv.org/html/2610.07250#bib.bib3)\); across tasks, it accumulates experience so that useful strategies carry over to later requests\([Chen et al\., 2026b](https://arxiv.org/html/2610.07250#bib.bib5)\)\. As the harness evolves, it learns to use the generator better\. The generator itself, however, does not change: the acquired gains live in external guidance that must be retrieved and applied again on every subsequent request\.

Keeping these gains outside the generator has two costs\. First, as memories and skills accumulate, relevant guidance becomes harder to retrieve\. Second, reapplying this guidance to every request incurs repeated reasoning, verification, and image generation\. Internalizing established gains could reduce both costs: a one\-time offline training cost is amortized over future requests, and obsolete strategies can be discarded so the harness can target the updated generator’s remaining limitations\. This raises our central question: can gains discovered through harness evolution be internalized into the generator, enabling it to retain these improvements without external guidance at inference time?

Answering this question requires an interface between harness experience and generator learning\. Although agents represent experience through different memories, skills, and reasoning traces, all ultimately condition the T2I generator on a text prompt\. We therefore retain, for each completed task, the original query and the harness\-generated prompt that yielded the selected image\. This captures prompt\-mediated gains, but not gains from verification or stochastic sample selection\. Prior work uses visual experience to improve the agent policy that constructs generator inputs\([Chen et al\., 2026a](https://arxiv.org/html/2610.07250#bib.bib6)\), leaving the harness responsible for realizing these gains at inference time\. In contrast, we use the resulting query–prompt pairs to train the generator itself, making the gains available from the original query alone\.

To perform this transfer, we introduce*Diffusion On\-Policy Context Distillation \(D\-OPCD\)*, a self\-distillation algorithm for diffusion models\. A teacher conditioned on both the original query and the harness\-produced prompt supervises a student that receives only the original query, at states sampled from the student’s own denoising trajectories\. We formulate harness\-to\-generator transfer as on\-policy context distillation\([Ye et al\., 2026](https://arxiv.org/html/2610.07250#bib.bib10)\)for diffusion models, with the evolving harness providing privileged textual context\. Unlike prior on\-policy diffusion self\-distillation\([Jiang et al\., 2026a](https://arxiv.org/html/2610.07250#bib.bib9)\), which obtains privileged information from target or reference images, D\-OPCD distills the gains elicited by the harness\. It thereby internalizes prompt\-mediated harness gains into the generator’s weights, making them available without the harness at inference time\.

We study this process with Auto Skill Evolver \(ASE\), a test\-time learning mechanism that turns completed tasks into reusable prompt guidance; the resulting query–prompt pairs train D\-OPCD\. Distilling these pairs improves the generator’s direct\-generation performance across all four benchmarks\. With the skill\-free harness reattached, the updated generator outperforms its base\-generator counterpart and even slightly exceeds the original skill\-equipped agent on average, indicating that its learned skills have been internalized\. Crucially, this establishes an iterative improvement cycle: after the generator internalizes the learned skills, resetting those now\-redundant skills allows ASE to evolve again, yielding further improvements\.

In summary, our main contributions are as follows:

- •We introduce D\-OPCD, an on\-policy context distillation algorithm that uses harness\-produced prompts as privileged teacher context to internalize prompt\-mediated gains into diffusion model weights, making part of the harness’s benefit available from the original query alone\.
- •We develop Auto Skill Evolver \(ASE\), a test\-time learning mechanism that turns task experience into reusable skills and provides query–prompt pairs for D\-OPCD\. Experiments across four benchmarks demonstrate improved direct generation after internalization\.
- •We formulate harness–model co\-evolution for T2I agents by coupling harness adaptation with generator learning\. Internalizing accumulated harness gains into the generator allows the harness to shed its saturated skills and resume evolving around the updated model\.

## 2Related Work

### 2\.1Agentic Text\-to\-Image Generation

The quality of a generated image depends in part on how a user’s request is translated into instructions for the generator\. Prompt adaptation improves this interface by rewriting requests into prompts better suited to a fixed T2I model\([Hao et al\., 2023](https://arxiv.org/html/2610.07250#bib.bib24)\)\. For more complex requests, image\-generation agents combine planning, generation, visual inspection, and prompt revision, using feedback from earlier attempts to address missing objects or unsatisfied constraints\([Wang et al\., 2025](https://arxiv.org/html/2610.07250#bib.bib2);[Jiang et al\., 2026b](https://arxiv.org/html/2610.07250#bib.bib1);[Kovalev et al\., 2025](https://arxiv.org/html/2610.07250#bib.bib25);[Zhang et al\., 2026b](https://arxiv.org/html/2610.07250#bib.bib4)\)\. This process can also extend beyond a single task: GEMS equips the agent with memory and skills, MemoGen reuses experience from earlier requests, and GenEvolve learns from generation trajectories to improve the agent’s decisions\([He et al\., 2026](https://arxiv.org/html/2610.07250#bib.bib3);[Chen et al\., 2026b](https://arxiv.org/html/2610.07250#bib.bib5);[Chen et al\., 2026a](https://arxiv.org/html/2610.07250#bib.bib6)\)\. These approaches show how better prompt construction and accumulated experience can improve the use of an image generator over time\.

### 2\.2Training Diffusion Models

Updating a diffusion generator’s weights provides another route to improving its outputs\. Supervised fine\-tuning can teach new visual concepts from a small set of examples, as in DreamBooth, or support continual customization across concepts\([Ruiz et al\., 2023](https://arxiv.org/html/2610.07250#bib.bib26);[Smith et al\., 2023](https://arxiv.org/html/2610.07250#bib.bib8)\)\. For qualities that are harder to specify with target images, training can instead use evaluative signals: ImageReward supplies preference\-based feedback for ReFL, DDPO optimizes rewards over denoising trajectories, DRaFT backpropagates differentiable rewards through sampling, and Diffusion\-DPO learns from preferred image pairs\([Xu et al\., 2023](https://arxiv.org/html/2610.07250#bib.bib11);[Black et al\., 2024](https://arxiv.org/html/2610.07250#bib.bib27);[Clark et al\., 2024](https://arxiv.org/html/2610.07250#bib.bib28);[Wallace et al\., 2024](https://arxiv.org/html/2610.07250#bib.bib17)\)\. Recent methods extend online reward optimization to flow\-matching models and use teacher predictions along a student’s own sampling trajectory to improve diffusion or flow generators\([Liu et al\., 2025](https://arxiv.org/html/2610.07250#bib.bib18);[Fang et al\., 2026](https://arxiv.org/html/2610.07250#bib.bib19);[Jiang et al\., 2026a](https://arxiv.org/html/2610.07250#bib.bib9)\)\. This literature offers several ways to turn examples or feedback into persistent changes in generation behavior\.

### 2\.3Harness–Model Co\-Evolution

Experience\-driven agent improvement begins by recording what succeeded or failed and turning those observations into guidance for later tasks\. ExpeL extracts reusable lessons from trajectories, while MemSkill learns and revises skills for managing an agent’s memory\([Zhao et al\., 2024](https://arxiv.org/html/2610.07250#bib.bib16);[Zhang et al\., 2026a](https://arxiv.org/html/2610.07250#bib.bib15)\)\. Recent work connects this external adaptation to policy learning by evolving skills alongside reinforcement learning\([Xia et al\., 2026a](https://arxiv.org/html/2610.07250#bib.bib13);[Zhang et al\., 2026c](https://arxiv.org/html/2610.07250#bib.bib29)\), updating an agent’s harness and model across interactions\([Xia et al\., 2026b](https://arxiv.org/html/2610.07250#bib.bib14);[Karten et al\., 2026](https://arxiv.org/html/2610.07250#bib.bib7)\), or deriving hindsight skills that supervise the current policy\([Wu et al\., 2026](https://arxiv.org/html/2610.07250#bib.bib30)\)\. Across these settings, the current model shapes the experience available for the next update, while accumulated experience changes the model and the guidance surrounding it\. We study this interaction in T2I generation, where the evolving harness constructs prompts for a separate image generator\.

![Refer to caption](https://arxiv.org/html/2610.07250v1/method_overview.png)Figure 1:Method overview\.As tasks accumulate, a T2I harness learns reusable skills that improve its prompts while the generator remains fixed\. D\-OPCD conditions an EMA teacher on the original queryqiq\_\{i\}and the harness\-selected promptpip\_\{i\}, and trains a student conditioned only onqiq\_\{i\}to match the teacher’s predictions along the student’s own denoising trajectory\. The updated generator can generate fromqiq\_\{i\}alone, while the harness resets its skills to continue adapting around it\.

## 3Method

### 3\.1Harness Gains at the Generator Interface

In a training\-free T2I agent in which the harness may invoke the generator multiple times, the final performance gains arise mainly from two main channels: improving the conditioning supplied to a frozen diffusion model and using external tools to compensate for its limitations\. Harness components—including memory, skills, planning, verification, and reflection—may interact in harness\-specific ways to realize these gains\. We focus on the prompt\-mediated gain exposed at the generator interface: for a task with its original queryqiq\_\{i\}, each invocation of the diffusion model produces a prompt–image pair,

𝒢i=\{\(pi,j,Ii,j\)\}j=1Ji,Ii,j∼πθ\(⋅∣pi,j\)\\mathcal\{G\}\_\{i\}=\\\{\(p\_\{i,j\},I\_\{i,j\}\)\\\}\_\{j=1\}^\{J\_\{i\}\},\\qquad I\_\{i,j\}\\sim\\pi\_\{\\theta\}\(\\cdot\\mid p\_\{i,j\}\)\(1\)where eachpi,jp\_\{i,j\}records a generation condition constructed by the harness andJiJ\_\{i\}is the number of diffusion model generation made by the harness for taskii\. Letji⋆j\_\{i\}^\{\\star\}denote the call selected by the harness for submission, and define

\(pi,Ii\):=\(pi,ji⋆,Ii,ji⋆\)∈𝒢i\.\(p\_\{i\},I\_\{i\}\):=\(p\_\{i,j\_\{i\}^\{\\star\}\},I\_\{i,j\_\{i\}^\{\\star\}\}\)\\in\\mathcal\{G\}\_\{i\}\.\(2\)
Under a fixed generator and sampling protocol, the quality ofIiI\_\{i\}provides an observable proxy for the effectiveness ofpip\_\{i\}\. The selected pair therefore provides a design\-agnostic interface for internalization:pip\_\{i\}encodes the condition constructed by the harness, whileIiI\_\{i\}records its realized outcome\.

### 3\.2Diffusion On Policy Context Distillation

#### 3\.2\.1Harness Internalization as Context Transfer

The selected promptpip\_\{i\}is constructed and retained by the harness after reasoning, refinement, and verification\. It therefore serves as a task\-specific compilation of the harness’s prompt\-mediated experience, while the associated imageIiI\_\{i\}provides evidence that this condition is effective\. Our goal is not merely to reproduceIiI\_\{i\}frompip\_\{i\}, but to transfer the generation advantageencodedinpip\_\{i\}to the model when it receives only the original queryqiq\_\{i\}\.

Direct SFT cannot realize this transfer\. Training on\(pi,Ii\)\(p\_\{i\},I\_\{i\}\)leaves the harness knowledge in the input and does not teach query\-only generation, whereas training on\(qi,Ii\)\(q\_\{i\},I\_\{i\}\)discards this privileged signal and supervises image\-derived rather than student\-visited states, creating an off\-policy mismatch\. We therefore introduce*Diffusion On\-Policy Context Distillation*\(D\-OPCD\), which the teacher has access to privileged context that is hidden from the student\([Ye et al\., 2026](https://arxiv.org/html/2610.07250#bib.bib10)\)\.

Specifically, D\-OPCD uses a teacher diffusion model conditioned on bothqiq\_\{i\}andpip\_\{i\}to supervise a student model conditioned only onqiq\_\{i\}, with both evaluated along the student’s own denoising trajectory, following the on\-policy diffusion construction\([Jiang et al\., 2026a](https://arxiv.org/html/2610.07250#bib.bib9)\)\. This provides a training signal without additional evaluator queries and directly targets the transfer from harness\-assisted generation to generation under the original query\.

#### 3\.2\.2On\-Policy Distillation Objective

##### Harness version and context dataset\.

Letℋr\\mathcal\{H\}\_\{r\}denote a snapshot of the harness, whererrindexes its version\. The harness may evolve through self\-evolution \(Section[3\.3](https://arxiv.org/html/2610.07250#S3.SS3)\) or human tuning, producing successive versions\. A versionℋr\\mathcal\{H\}\_\{r\}is admitted for internalization only if it improves over query\-only generation on the held\-out task set𝒬hold\\mathcal\{Q\}\_\{\\mathrm\{hold\}\}\(see the data protocol in Section[3\.4](https://arxiv.org/html/2610.07250#S3.SS4)\)\. Letπθ\\pi\_\{\\theta\}denote the generator on whichℋr\\mathcal\{H\}\_\{r\}operates\. This generator stays frozen whileℋr\\mathcal\{H\}\_\{r\}produces data and is modified only by the D\-OPCD update described below\.

Runningℋr\\mathcal\{H\}\_\{r\}withπθ\\pi\_\{\\theta\}on𝒬train\\mathcal\{Q\}\_\{\\mathrm\{train\}\}produces completed tasks\. Each task has a submitted pair\(pi,Ii\)\(p\_\{i\},I\_\{i\}\), defined in Equation[2](https://arxiv.org/html/2610.07250#S3.E2)\. Recall that the harness selectspip\_\{i\}itself:pip\_\{i\}is the final generation condition whose realized image passed the harness’s internal verification\. We retainNrN\_\{r\}completed\-task records to form the context dataset forℋr\\mathcal\{H\}\_\{r\}:

𝒟ctx\(r\)=\{\(qi,pi,Ii\)\}i=1Nr\.\\mathcal\{D\}\_\{\\mathrm\{ctx\}\}^\{\(r\)\}=\\\{\(q\_\{i\},p\_\{i\},I\_\{i\}\)\\\}\_\{i=1\}^\{N\_\{r\}\}\.\(3\)The imageIiI\_\{i\}is used only for record selection, where it filters out tasks whose submitted outputs fail the selection criteria\. D\-OPCD only usespip\_\{i\}as the sole privileged context\.

##### Asymmetric conditioning\.

The text\-conditioning encoderFTF\_\{\\mathrm\{T\}\}maps text into the generator’s conditioning space\. We define two branches:

cis=FT​\(qi\),cit=FT​\(\[qi;pi\]\),c\_\{i\}^\{s\}=F\_\{\\mathrm\{T\}\}\(q\_\{i\}\),\\qquad c\_\{i\}^\{t\}=F\_\{\\mathrm\{T\}\}\(\[q\_\{i\};p\_\{i\}\]\),\(4\)where\[qi;pi\]\[q\_\{i\};p\_\{i\}\]denotes a structured text condition consisting of the original query followed by the harness\-selected prompt\. The student receives only the original user condition\. The teacher receives the same condition together with the context constructed by the harness\.

##### On\-policy rollout\.

Following D\-OPSD\([Jiang et al\., 2026a](https://arxiv.org/html/2610.07250#bib.bib9)\), letvθ​\(xt,t,c\)v\_\{\\theta\}\(x\_\{t\},t,c\)denote the velocity field of the student generator, initialized fromθ\\theta\. The teachervθ¯v\_\{\\bar\{\\theta\}\}is initialized from the same parameters and is updated as an exponential moving average of the student\. Let0=t1<⋯<tK<tK\+1=10=t\_\{1\}<\\cdots<t\_\{K\}<t\_\{K\+1\}=1be the rollout schedule, and letΦ\\Phibe the diffusion solver used at deployment\. Starting fromxi,t1s=ϵx\_\{i,t\_\{1\}\}^\{s\}=\\epsilon, whereϵ∼𝒩⁡\(0,𝐈\)\\epsilon\\sim\\mathcal\{N\}\(0,\\mathbf\{I\}\), the student produces its own rollout undercisc\_\{i\}^\{s\}:

xi,tk\+1s=Φ\(xi,tks,tk,tk\+1,vθ\(⋅,⋅,cis\)\),k=1,…,K\.x\_\{i,t\_\{k\+1\}\}^\{s\}=\\Phi\\\!\\left\(x\_\{i,t\_\{k\}\}^\{s\},t\_\{k\},t\_\{k\+1\},v\_\{\\theta\}\(\\cdot,\\cdot,c\_\{i\}^\{s\}\)\\right\),\\qquad k=1,\\ldots,K\.\(5\)The visited states are treated as constants, so gradients do not propagate through the rollout\. At each visited state, the student and the teacher predict velocities under their respective conditions:

ui,ks=vθ​\(xi,tks,tk,cis\),ui,kt=vθ¯​\(xi,tks,tk,cit\)\.u\_\{i,k\}^\{s\}=v\_\{\\theta\}\(x\_\{i,t\_\{k\}\}^\{s\},t\_\{k\},c\_\{i\}^\{s\}\),\\qquad u\_\{i,k\}^\{t\}=v\_\{\\bar\{\\theta\}\}\(x\_\{i,t\_\{k\}\}^\{s\},t\_\{k\},c\_\{i\}^\{t\}\)\.\(6\)

##### Objective\.

We train the student to match the privileged\-context teacher on these shared on\-policy states:

ℒD\-OPCD=𝔼\(qi,pi\)∼𝒟ctx\(r\),ϵ∼𝒩⁡\(0,𝐈\)​\[1K​∑k=1K‖ui,ks−sg⁡\(ui,kt\)‖22\],\\mathcal\{L\}\_\{\\text\{D\-OPCD\}\}=\\mathbb\{E\}\_\{\(q\_\{i\},p\_\{i\}\)\\sim\\mathcal\{D\}\_\{\\mathrm\{ctx\}\}^\{\(r\)\},\\;\\epsilon\\sim\\mathcal\{N\}\(0,\\mathbf\{I\}\)\}\\left\[\\frac\{1\}\{K\}\\sum\_\{k=1\}^\{K\}\\left\\\|u\_\{i,k\}^\{s\}\-\\operatorname\{sg\}\\\!\\left\(u\_\{i,k\}^\{t\}\\right\)\\right\\\|\_\{2\}^\{2\}\\right\],\(7\)
wheresg⁡\(⋅\)\\operatorname\{sg\}\(\\cdot\)denotes stop\-gradient\. This objective combines OPCD’s context asymmetry with D\-OPSD’s diffusion\-level on\-policy alignment: both branches shareqiq\_\{i\},pip\_\{i\}supplies additional teacher\-only advantage information, and their predictions are compared on states visited by the student\.

### 3\.3Self\-Evolving Harness via Skill Extraction

To demonstrate a harness whose prompt\-mediated gain improves with experience, we construct the*Auto Skill Evolver*\(ASE\), a hierarchical test\-time learning framework that distills agent execution trajectories into reusable skills\. Figure[2](https://arxiv.org/html/2610.07250#S3.F2)summarizes the harness evolution process\.

![Refer to caption](https://arxiv.org/html/2610.07250v1/harness_evolution.png)Figure 2:Harness evolution with ASE, from task episodes through insights to skill updates\.ASE organizes experience into three levels: 1\) An*episode*is the agent’s complete execution trajectory on a single task together with task\-level feedback returned by an external evaluator after image submission; 2\) An*insight*, extracted from episodes, records what the agent discovered during execution that could have been provided in advance; 3\) A*skill*consolidates multiple insights into reusable guidance for constructing the first generation prompt; within the scope of this paper, skills do not invoke tools\. At the start of each task, the agent loads the current skill library𝒮\\mathcal\{S\}and selects applicable skills to construct its first generation prompt, applying guidance learned from earlier tasks to the current request\. Verification feedback and task\-local history guide subsequent prompt refinement\. Further details of the agent harness and ASE implementation are given in Appendix[C](https://arxiv.org/html/2610.07250#A3)\.

##### Insight extraction\.

Tasks arrive as a stream and are processed in batches ofMMtasks\. During harness evolution, the agent may retrieve relevant insights from the pool𝒫\\mathcal\{P\}after failed verification to guide prompt refinement\. The retrieved insights and the task outcome are recorded together, providing evidence about their usefulness across tasks\. After each batch, an insight manager separately retrieves relevant insights from𝒫\\mathcal\{P\}to match against the completed episodes and may create, support, revise, contradict, merge, or archive insights\. An insight becomes*mature*once its support count reachesτ\\tauand exceeds its contradiction count\.

##### Skill update\.

Once𝒫\\mathcal\{P\}contains at leastBBmature insights not yet absorbed into the library, a skill manager examines them together with𝒮\\mathcal\{S\}and may leave the library unchanged or propose to create, revise, split, merge, or retire skills\. Invalid plans are retried with validation feedback; if the staged library exceeds size or length limits, the manager reorganizes it before acceptance\. When a proposal is accepted, its changes are committed and the insights it absorbs are marked as*reviewed*, excluding them from subsequent updates so that each update draws only on new evidence\.

##### Harness versions\.

Frozen skill\-only replay and evaluation disable cross\-task insight retrieval and ASE updates\. Accumulated experience therefore reaches the agent only through the committed skill library, so each committed update defines a new harness version\. We denote byℋr\\mathcal\{H\}\_\{r\}the harness equipped with therr\-th skill library𝒮r\\mathcal\{S\}\_\{r\}\. Among the resulting versions, the one to internalize with D\-OPCD is selected on the held\-out task set𝒬hold\\mathcal\{Q\}\_\{\\mathrm\{hold\}\}\.

### 3\.4Harness–Generator Co\-Evolution

Harness evolution improves the agent without training, but the harness has limited capacity: each added skill consumes context and makes it harder for the agent to choose the right one at inference time\. D\-OPCD relieves this limit by internalizing part of the harness knowledge into the generator weights\. After the generator update, we clear the learned skills to examine whether the harness can acquire new guidance around the updated generator\.

##### Data protocol\.

For each benchmark, we split the tasks into three disjoint sets,𝒬=𝒬train⊔𝒬hold⊔𝒬eval\\mathcal\{Q\}=\\mathcal\{Q\}\_\{\\mathrm\{train\}\}\\sqcup\\mathcal\{Q\}\_\{\\mathrm\{hold\}\}\\sqcup\\mathcal\{Q\}\_\{\\mathrm\{eval\}\}\.𝒬train\\mathcal\{Q\}\_\{\\mathrm\{train\}\}is used for harness evolution and D\-OPCD training;𝒬hold\\mathcal\{Q\}\_\{\\mathrm\{hold\}\}only for selecting harness versions and training checkpoints; and𝒬eval\\mathcal\{Q\}\_\{\\mathrm\{eval\}\}only for reporting final results\.

##### Harness evolution and version selection\.

With the current generatorπθ\\pi\_\{\\theta\}fixed, the agent processes𝒬train\\mathcal\{Q\}\_\{\\mathrm\{train\}\}as a stream and evolves its skill library with ASE \(Section[3\.3](https://arxiv.org/html/2610.07250#S3.SS3)\), producing harness versions\{ℋr\}\\\{\\mathcal\{H\}\_\{r\}\\\}\. We evaluate each version on𝒬hold\\mathcal\{Q\}\_\{\\mathrm\{hold\}\}and select the best\-performing versionℋr⋆\\mathcal\{H\}\_\{r^\{\\star\}\}using the external evaluator’s aggregate scoreJJ\.

r⋆=arg⁡maxr⁡J⁡\(ℋr,πθ,𝒬hold\)\.r^\{\\star\}=\\arg\\max\_\{r\}\\ J\\big\(\\mathcal\{H\}\_\{r\};\\ \\pi\_\{\\theta\},\\ \\mathcal\{Q\}\_\{\\mathrm\{hold\}\}\\big\)\.\(8\)

##### Replay and internalization\.

We freeze the skill library ofℋr⋆\\mathcal\{H\}\_\{r^\{\\star\}\}and replay it on𝒬train\\mathcal\{Q\}\_\{\\mathrm\{train\}\}withπθ\\pi\_\{\\theta\}\. The resulting records form𝒟ctx\(r⋆\)\\mathcal\{D\}\_\{\\mathrm\{ctx\}\}^\{\(r^\{\\star\}\)\}in Equation[3](https://arxiv.org/html/2610.07250#S3.E3), and D\-OPCD uses this dataset to update the generator toπθ′\\pi\_\{\\theta^\{\\prime\}\}\. We replay rather than reuse the streaming records so that all contexts come from a single harness version operating on the generator from which D\-OPCD is initialized\.

##### Generator adoption and harness reset\.

After D\-OPCD, we adoptπθ′\\pi\_\{\\theta^\{\\prime\}\}as the new generator\. To measure how much of the harness gain has been transferred into the weights, we define the signed internalization gap

δ\(r⋆\)=J⁡\(ℋ∅,πθ′\)−J⁡\(ℋr⋆,πθ\)\.\\delta^\{\(r^\{\\star\}\)\}=J\\big\(\\mathcal\{H\}\_\{\\emptyset\};\\pi\_\{\\theta^\{\\prime\}\}\\big\)\-J\\big\(\\mathcal\{H\}\_\{r^\{\\star\}\};\\pi\_\{\\theta\}\\big\)\.\(9\)Both scores are measured on𝒬eval\\mathcal\{Q\}\_\{\\mathrm\{eval\}\}\. A value near zero means that the updated generator without learned skills matches the original skill\-equipped agent; a positive value favors the updated generator, while a negative value indicates a remaining deficit\. We clear the skill library and its associated episode and insight records, retain the skill\-free harnessℋ∅\\mathcal\{H\}\_\{\\emptyset\}, and evolve new harness versions againstπθ′\\pi\_\{\\theta^\{\\prime\}\}\. We denote byℋr†\\mathcal\{H\}\_\{r^\{\\dagger\}\}a version selected on𝒬hold\\mathcal\{Q\}\_\{\\mathrm\{hold\}\}underπθ′\\pi\_\{\\theta^\{\\prime\}\}after this renewed harness evolution\.

## 4Experiments

### 4\.1Experimental Setup

##### Tasks and benchmarks\.

We evaluate on GenEval\([Ghosh et al\., 2023](https://arxiv.org/html/2610.07250#bib.bib12)\), GenEval2\([Kamath et al\., 2025](https://arxiv.org/html/2610.07250#bib.bib20)\), WISE\([Niu et al\., 2025](https://arxiv.org/html/2610.07250#bib.bib21)\), and R2I\-Bench\([Chen et al\., 2025](https://arxiv.org/html/2610.07250#bib.bib22)\), which cover compositional, knowledge\-intensive, and reasoning\-intensive generation; T2I\-CompBench\+\+\([Huang et al\., 2025](https://arxiv.org/html/2610.07250#bib.bib23)\)is reserved for out\-of\-distribution \(OOD\) evaluation\. Each benchmark forms a sequential task stream: the agent receives a query and obtains evaluator feedback after submitting its final image\. We use𝒬train\\mathcal\{Q\}\_\{\\mathrm\{train\}\},𝒬hold\\mathcal\{Q\}\_\{\\mathrm\{hold\}\}, and𝒬eval\\mathcal\{Q\}\_\{\\mathrm\{eval\}\}as defined in Section[3\.4](https://arxiv.org/html/2610.07250#S3.SS4), and report evaluation results on𝒬eval\\mathcal\{Q\}\_\{\\mathrm\{eval\}\}\. The official evaluators provide GenEval Overall, GenEval2 Soft\-TIFA AM, WIScore, and R2I\-Score, respectively\. Training\-method comparisons use direct generation fromqiq\_\{i\}; the main results also evaluate the base and D\-OPCD\-updated generators with skill\-free and evolved harnesses\. Appendix[E\.1](https://arxiv.org/html/2610.07250#A5.SS1)gives split sizes, and Appendix[E](https://arxiv.org/html/2610.07250#A5)details the evaluation protocols\.

##### Agent and harness evolution\.

We use Z\-Image\-Turbo as the initial T2I generator\. The agent starts from a GEMS\-based scaffold\([He et al\., 2026](https://arxiv.org/html/2610.07250#bib.bib3)\)with an empty skill library \(the skill\-free harness\) and is equipped with the Auto Skill Evolver \(ASE\) described in Section[3\.3](https://arxiv.org/html/2610.07250#S3.SS3)\. With the generator fixed, it processes𝒬train\\mathcal\{Q\}\_\{\\mathrm\{train\}\}as a task stream; ASE extracts insights from completed episodes and consolidates mature insights into skills, producing successive harness versions\. Appendix[C\.1](https://arxiv.org/html/2610.07250#A3.SS1)describes the scaffold design, and Appendix[C\.4](https://arxiv.org/html/2610.07250#A3.SS4)reports the benchmark\-specific ASE settings\.

##### Compared training methods\.

For a matched comparison, all methods use the same prompt\-changed task records replayed from the selected skill\-equipped harnessℋr⋆\\mathcal\{H\}\_\{r^\{\\star\}\}on𝒬train\\mathcal\{Q\}\_\{\\mathrm\{train\}\}\. They start from the same Z\-Image\-Turbo weights and train LoRA adapters with the same optimization budget\. D\-OPCD uses each\(qi,pi\)\(q\_\{i\},p\_\{i\}\)pair: an EMA teacher conditioned on\[qi;pi\]\[q\_\{i\};p\_\{i\}\]supervises a student conditioned only onqiq\_\{i\}along the student’s own sampling trajectory, withpip\_\{i\}hidden at inference\. Vanilla SFT fits the agent\-submitted imageIiI\_\{i\}fromqiq\_\{i\}with a standard flow\-matching objective\. Diffusion\-DPO\([Wallace et al\., 2024](https://arxiv.org/html/2610.07250#bib.bib17)\)treatsIiI\_\{i\}as chosen and a same\-seed image generated directly fromqiq\_\{i\}as rejected, training both underqiq\_\{i\}\. D\-OPSD\([Jiang et al\., 2026a](https://arxiv.org/html/2610.07250#bib.bib9)\)uses an image\-conditioned teacher based on\(qi,Ii\)\(q\_\{i\},I\_\{i\}\)to supervise aqiq\_\{i\}\-conditioned student along its own trajectory\. Appendix[D\.1](https://arxiv.org/html/2610.07250#A4.SS1)details the record selection; Appendix[D](https://arxiv.org/html/2610.07250#A4)gives training objectives and settings\.

### 4\.2Main Results

Table 1:Base and D\-OPCD generator scores under direct generation, a skill\-free harness, and evolved skills\. ParenthesizedΔ\\Deltavalues compare base\+\+rows with base direct generation and D\-OPCD rows with the corresponding rows in the base\-generator block above, in the same order\.##### Harness gains at the generator interface\.

To assess the gains provided by the agent harness, we evaluate the base generatorπθ\\pi\_\{\\theta\}on𝒬eval\\mathcal\{Q\}\_\{\\mathrm\{eval\}\}alone, with the skill\-free harnessℋ∅\\mathcal\{H\}\_\{\\emptyset\}, and with the selected skill\-equipped harnessℋr⋆\\mathcal\{H\}\_\{r^\{\\star\}\}evolved by ASE\. Both harness configurations improve performance over direct generation on all four benchmarks \(Table[1](https://arxiv.org/html/2610.07250#S4.T1)\)\. The skill\-free harness raises the four\-benchmark average from 60\.52 to 79\.28 \(\+18\.76 points\), including gains from 42\.86 to 78\.00 on WISE and from 42\.87 to 68\.12 on R2I\-Bench\. ASE\-evolved skills raise the average further to 81\.83 \(\+2\.55 points overℋ∅\\mathcal\{H\}\_\{\\emptyset\}\), with additional gains on every benchmark, ranging from \+1\.29 points on R2I\-Bench to \+4\.12 on GenEval\. The prompts selected byℋr⋆\\mathcal\{H\}\_\{r^\{\\star\}\}then provide the privileged contexts for D\-OPCD\.

##### Internalizing harness gains\.

To test whether the harness’s gains can be internalized, we train the generator on\(qi,pi\)\(q\_\{i\},p\_\{i\}\)pairs fromℋr⋆\\mathcal\{H\}\_\{r^\{\\star\}\}with D\-OPCD and evaluateπθ′\\pi\_\{\\theta^\{\\prime\}\}directly fromqiq\_\{i\}, without a harness\. It outperforms the base generator on all four benchmarks, raising the average score from 60\.52 to 65\.09 \(Table[1](https://arxiv.org/html/2610.07250#S4.T1)\)\. Some of the harness benefit therefore persists in the model weights\.

##### Harness benefits after internalization\.

To test whether internalization reduces reliance on the original evolved skills, we attach the skill\-free harnessℋ∅\\mathcal\{H\}\_\{\\emptyset\}toπθ′\\pi\_\{\\theta^\{\\prime\}\}and compare it with the base generator underℋ∅\\mathcal\{H\}\_\{\\emptyset\}andℋr⋆\\mathcal\{H\}\_\{r^\{\\star\}\}\. The updated skill\-free agent outperforms its base\-generator counterpart on all four benchmarks \(Table[1](https://arxiv.org/html/2610.07250#S4.T1)\)\. It scores 82\.33 on average and slightly exceeds the original skill\-equipped agent, yielding a signed internalization gap ofδ\(r⋆\)=\+0\.50\\delta^\{\(r^\{\\star\}\)\}=\+0\.50\. On R2I\-Bench, it is also ahead, withδ\(r⋆\)=\+4\.27\\delta^\{\(r^\{\\star\}\)\}=\+4\.27\. Thus, much of the earlier skill benefit can be retained after removing those skills, although the gap varies across benchmarks\.

##### Continued co\-evolution after model update\.

After adoptingπθ′\\pi\_\{\\theta^\{\\prime\}\}, we reset the learned skills and rerun ASE to obtainℋr†\\mathcal\{H\}\_\{r^\{\\dagger\}\}\. The renewed skills improve the updated agent on three benchmarks and raise its average above both the updated skill\-free agent and the original skill\-equipped agent \(Table[1](https://arxiv.org/html/2610.07250#S4.T1)\)\. R2I\-Bench instead declines after skill evolution\. The D\-OPCD generator with the skill\-free harness already scores highly, which may leave less room for additional gains\. These results illustrate how D\-OPCD unlocks a harness–model co\-evolution mechanism, with renewed skills building on the updated generator\. Matched image examples are provided in Appendix[A](https://arxiv.org/html/2610.07250#A1)\.

### 4\.3Ablation Studies and Further Analysis

#### 4\.3\.1Comparison with Alternative Training Methods

To compare methods for internalizing agent experience, we train D\-OPCD and three baselines on matched records collected by the evolved harness and evaluate direct generation from the original queries \(Table[2](https://arxiv.org/html/2610.07250#S4.T2)\)\. All four improve the base generator’s four\-benchmark average, with gains ranging from 1\.66 to 4\.57 points\. D\-OPCD achieves the highest mean of 65\.09, 1\.71 points above Vanilla SFT, and leads on all four benchmarks\. On WISE, D\-OPCD improves over the base generator by 7\.14 points\. Overall, these results suggest that D\-OPCD has an advantage over the tested baselines in internalizing agent experience into the generator’s model weights\.

Table 2:Direct\-generation comparison on𝒬eval\\mathcal\{Q\}\_\{\\mathrm\{eval\}\}\. Updated models train on matched records fromℋr⋆\\mathcal\{H\}\_\{r^\{\\star\}\}: D\-OPCD uses\(qi,pi\)\(q\_\{i\},p\_\{i\}\), while baselines use the agent imageIiI\_\{i\}\. ParenthesizedΔ\\Deltavalues are differences from the base generator; Avg\. is the unweighted mean across four benchmarks\.
#### 4\.3\.2Ablation of Teacher Conditioning

With the student conditioned onqiq\_\{i\}, we compare three teacher inputs: matched\[qi;pi\]\[q\_\{i\};p\_\{i\}\],pip\_\{i\}alone, and shuffled\[qi;pσ⁡\(i\)\]\[q\_\{i\};p\_\{\\sigma\(i\)\}\]using another task’s prompt\. The matched input averages 65\.09, ahead ofpip\_\{i\}alone \(64\.63\) and shuffled prompts \(61\.05\)\. It leads on three benchmarks, whilepip\_\{i\}alone reaches the highest R2I\-Bench score of 47\.32\. Shuffling reduces scores on all four benchmarks, underscoring the value of task\-matched teacher prompts \(Table[3](https://arxiv.org/html/2610.07250#S4.T3)\)\.

Table 3:D\-OPCD performance with different teacher inputs on𝒬eval\\mathcal\{Q\}\_\{\\mathrm\{eval\}\}\. The teacher receives the original query with its matched harness prompt \(\[qi;pi\]\[q\_\{i\};p\_\{i\}\]\), the harness prompt alone \(pip\_\{i\}only\), or the query with a prompt from another task \(Shuffled\)\. The student always receivesqiq\_\{i\}\.
#### 4\.3\.3Context Properties and Model Internalization

Table[1](https://arxiv.org/html/2610.07250#S4.T1)shows that training on contexts replayed from the selected skill\-equipped harnessℋr⋆\\mathcal\{H\}\_\{r^\{\\star\}\}improves direct generation from the original query on all four benchmarks\. To explore how context properties relate to model internalization, we examine two observable properties of the collected\(qi,pi\)\(q\_\{i\},p\_\{i\}\)pairs: the submitted image’s advantage over direct\-query generation and variation in prompt wording\. We compare these data\-level properties with direct\-generation scores after training\.

##### Context design\.

We vary the default prompt\-changed replay set in two ways\.1\) Improvement filter\.We retain a pair only when the agent’s internal verifier gives the submitted image a strictly higher pass count than the direct\-query image\.2\) Diversity adapter\.We change the agent’s prompt\-sampling policy to reduce repeated prompt patterns, then collect its prompt\-changed pairs\. Appendix[E\.3](https://arxiv.org/html/2610.07250#A5.SS3)gives the construction details for both variants\.

##### Results and analysis\.

The improvement\-filtered pairs show larger observed image\-score gaps on their retained tasks, while the diversity adapter increases character\-level variation in the collected prompts \(Appendix[E\.3](https://arxiv.org/html/2610.07250#A5.SS3)\)\. However, downstream gains are mixed: both variants improve individual benchmarks, but the filter leaves the four\-benchmark mean nearly unchanged and the adapter lowers it \(Table[4](https://arxiv.org/html/2610.07250#S4.T4)\)\. Filtering also changes the training set’s size and composition, so its effect cannot be attributed solely to task improvement\. A prompt that is more useful during generation is therefore not necessarily easier for the query\-only student to internalize when used as context\. Identifying which harness\-produced contexts best support model internalization remains an open question\.

Table 4:D\-OPCD direct\-generation test scores after training on three context sets\. Pair counts follow GenEval/GenEval2/WISE/R2I\-Bench order; Avg\. is the four\-benchmark mean before rounding\.

#### 4\.3\.4Out\-of\-Distribution Generalization

We evaluate the base generator and four source\-specific D\-OPCD checkpoints through direct generation without a harness on seven T2I\-CompBench\+\+ categories reserved for out\-of\-distribution testing\([Huang et al\., 2025](https://arxiv.org/html/2610.07250#bib.bib23)\)\. Color scores rise from 77\.45 for the base generator to 81\.61 and 81\.44 for the GenEval\- and GenEval2\-trained checkpoints, respectively\. Their 2D\-Spatial scores likewise reach 36\.19 and 37\.33, up from 32\.32 for the base \(Table[5](https://arxiv.org/html/2610.07250#S4.T5)\)\. The WISE\- and R2I\-Bench\-trained checkpoints remain close to the base overall\. These results show good out\-of\-distribution performance, especially with GenEval and GenEval2 as training sources\.

Table 5:Out\-of\-distribution category scores on T2I\-CompBench\+\+ for the base generator and four D\-OPCD checkpoints\. The suffix in each D\-OPCD method name identifies its training benchmark\.

## 5Conclusion

We introduce D\-OPCD to internalize prompt\-mediated gains from an evolving T2I harness into the generator’s weights, allowing the model to retain part of the harness’s benefit from the original query alone\. With ASE turning task experience into reusable skills, D\-OPCD improves direct generation across all four benchmarks\. The updated generator with a skill\-free harness slightly exceeds the original skill\-equipped agent on average, and resetting the old skills allows ASE to evolve again and further improve performance\. We hope this work guides the development of T2I systems in which accumulated experience becomes lasting generator capability, creating room for the harness and model to keep improving each other through continual co\-evolution\.

## References

- K\. Black, M\. Janner, Y\. Du, I\. Kostrikov, and S\. LevineTraining diffusion models with reinforcement learning\.InInternational Conference on Learning Representations,B\. Kim, Y\. Yue, S\. Chaudhuri, K\. Fragkiadaki, M\. Khan, and Y\. Sun \(Eds\.\),Vol\.2024,pp\. 4965–4987\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2024/file/14f75513f0f1ca01de1e826b52e6b840-Paper-Conference.pdf)Cited by:[§2\.2](https://arxiv.org/html/2610.07250#S2.SS2.p1.1)\.
- Chenet al\.\(2025\)K\. Chen, Z\. Lin, Z\. Xu, Y\. Shen, Y\. Yao, J\. Rimchala, J\. Zhang, and L\. HuangR2I\-bench: benchmarking reasoning\-driven text\-to\-image generation\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),Suzhou, China,pp\. 12595–12630\.External Links:[Link](https://aclanthology.org/2025.emnlp-main.636/),[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.636),ISBN 979\-8\-89176\-332\-6Cited by:[§E\.4](https://arxiv.org/html/2610.07250#A5.SS4.p1.1),[§4\.1](https://arxiv.org/html/2610.07250#S4.SS1.SSS0.Px1.p1.1)\.
- Chenet al\.\(2026a\)S\. Chen, Z\. Xing, T\. Ye, X\. Geng, Y\. Lin, J\. Lai, X\. He, F\. Zhai, J\. Gao, and L\. ZhuGenEvolve: self\-evolving image generation agents via tool\-orchestrated visual experience distillation\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2605.21605),[Link](https://arxiv.org/abs/2605.21605)Cited by:[§1](https://arxiv.org/html/2610.07250#S1.p3.1),[§2\.1](https://arxiv.org/html/2610.07250#S2.SS1.p1.1)\.
- Chenet al\.\(2026b\)W\. Chen, K\. Yu, B\. Tian, J\. Song, S\. Liang, H\. Jia, K\. Cheng, H\. Li, K\. Yuan, L\. Wang, J\. Wu, S\. Lai, and Y\. YueMemoGen: can past experience improve future text\-to\-image generation?\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2606.03243),[Link](https://arxiv.org/abs/2606.03243)Cited by:[§1](https://arxiv.org/html/2610.07250#S1.p1.1),[§2\.1](https://arxiv.org/html/2610.07250#S2.SS1.p1.1)\.
- Clarket al\.\(2024\)K\. Clark, P\. Vicol, K\. Swersky, and D\. FleetDirectly fine\-tuning diffusion models on differentiable rewards\.InInternational Conference on Learning Representations,B\. Kim, Y\. Yue, S\. Chaudhuri, K\. Fragkiadaki, M\. Khan, and Y\. Sun \(Eds\.\),Vol\.2024,pp\. 4793–4822\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2024/file/13e8be77982beb73d7ed0bbf122f9f3c-Paper-Conference.pdf)Cited by:[§2\.2](https://arxiv.org/html/2610.07250#S2.SS2.p1.1)\.
- Fanget al\.\(2026\)Z\. Fang, W\. Huang, Y\. Zeng, Y\. Zhao, S\. Chen, K\. Feng, Y\. Lin, L\. Chen, Z\. Chen, S\. Cao, and F\. ZhaoFlow\-opd: on\-policy distillation for flow matching models\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2605.08063),[Link](https://arxiv.org/abs/2605.08063)Cited by:[§2\.2](https://arxiv.org/html/2610.07250#S2.SS2.p1.1)\.
- Friedman and Dieng \(2023\)D\. Friedman and A\. B\. DiengThe vendi score: a diversity evaluation metric for machine learning\.Transactions on Machine Learning Research\.External Links:ISSN 2835\-8856Cited by:[§E\.3](https://arxiv.org/html/2610.07250#A5.SS3.SSS0.Px4.p1.1)\.
- Ghoshet al\.\(2023\)D\. Ghosh, H\. Hajishirzi, and L\. SchmidtGenEval: an object\-focused framework for evaluating text\-to\-image alignment\.InAdvances in Neural Information Processing Systems 36,NeurIPS 2023,pp\. 52132–52152\.External Links:[Link](http://dx.doi.org/10.52202/075280-2270),[Document](https://dx.doi.org/10.52202/075280-2270)Cited by:[§E\.4](https://arxiv.org/html/2610.07250#A5.SS4.p1.1),[§4\.1](https://arxiv.org/html/2610.07250#S4.SS1.SSS0.Px1.p1.1)\.
- Haoet al\.\(2023\)Y\. Hao, Z\. Chi, L\. Dong, and F\. WeiOptimizing prompts for text\-to\-image generation\.InAdvances in Neural Information Processing Systems,A\. Oh, T\. Naumann, A\. Globerson, K\. Saenko, M\. Hardt, and S\. Levine \(Eds\.\),Vol\.36,pp\. 66923–66939\.External Links:[Document](https://dx.doi.org/10.52202/075280-2923),[Link](https://proceedings.neurips.cc/paper_files/paper/2023/file/d346d91999074dd8d6073d4c3b13733b-Paper-Conference.pdf)Cited by:[§2\.1](https://arxiv.org/html/2610.07250#S2.SS1.p1.1)\.
- Heet al\.\(2026\)Z\. He, S\. Huang, X\. Qu, Y\. Li, T\. Zhu, Y\. Cheng, and Y\. YangGEMS: agent\-native multimodal generation with memory and skills\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2603.28088),[Link](https://arxiv.org/abs/2603.28088)Cited by:[§C\.1](https://arxiv.org/html/2610.07250#A3.SS1.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2610.07250#S1.p1.1),[§2\.1](https://arxiv.org/html/2610.07250#S2.SS1.p1.1),[§4\.1](https://arxiv.org/html/2610.07250#S4.SS1.SSS0.Px2.p1.1)\.
- Huanget al\.\(2025\)K\. Huang, C\. Duan, K\. Sun, E\. Xie, Z\. Li, and X\. LiuT2I\-compbench\+\+: an enhanced and comprehensive benchmark for compositional text\-to\-image generation\.IEEE Transactions on Pattern Analysis and Machine Intelligence47\(5\),pp\. 3563–3579\.External Links:ISSN 1939\-3539,[Link](http://dx.doi.org/10.1109/TPAMI.2025.3531907),[Document](https://dx.doi.org/10.1109/tpami.2025.3531907)Cited by:[§E\.4](https://arxiv.org/html/2610.07250#A5.SS4.p2.1),[§4\.1](https://arxiv.org/html/2610.07250#S4.SS1.SSS0.Px1.p1.1),[§4\.3\.4](https://arxiv.org/html/2610.07250#S4.SS3.SSS4.p1.1)\.
- Jianget al\.\(2026a\)D\. Jiang, X\. Jin, D\. Liu, Z\. Wang, M\. Zheng, R\. Du, X\. Yang, Q\. Wu, Z\. Li, P\. Gao, H\. Yang, and S\. HoiD\-opsd: on\-policy self\-distillation for continuously tuning step\-distilled diffusion models\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2605.05204),[Link](https://arxiv.org/abs/2605.05204)Cited by:[§1](https://arxiv.org/html/2610.07250#S1.p4.1),[§2\.2](https://arxiv.org/html/2610.07250#S2.SS2.p1.1),[§3\.2\.1](https://arxiv.org/html/2610.07250#S3.SS2.SSS1.p3.1),[§3\.2\.2](https://arxiv.org/html/2610.07250#S3.SS2.SSS2.Px3.p1.1),[§4\.1](https://arxiv.org/html/2610.07250#S4.SS1.SSS0.Px3.p1.1),[Table 2](https://arxiv.org/html/2610.07250#S4.T2.2.5.1.1.1)\.
- Jianget al\.\(2026b\)K\. Jiang, Y\. Wang, J\. Zhou, P\. Li, Z\. Liu, C\. Xie, Z\. Chen, Y\. Zheng, and W\. ZhangGenAgent: scaling text\-to\-image generation via agentic multimodal reasoning\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2601.18543),[Link](https://arxiv.org/abs/2601.18543)Cited by:[§2\.1](https://arxiv.org/html/2610.07250#S2.SS1.p1.1)\.
- Kamathet al\.\(2025\)A\. Kamath, K\. Chang, R\. Krishna, L\. Zettlemoyer, Y\. Hu, and M\. GhazvininejadGenEval 2: addressing benchmark drift in text\-to\-image evaluation\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2512.16853),[Link](https://arxiv.org/abs/2512.16853)Cited by:[§E\.4](https://arxiv.org/html/2610.07250#A5.SS4.p1.1),[§4\.1](https://arxiv.org/html/2610.07250#S4.SS1.SSS0.Px1.p1.1)\.
- Kartenet al\.\(2026\)S\. Karten, J\. Zhang, T\. Upaa, R\. Feng, W\. Li, C\. Shi, C\. Jin, and K\. VodrahalliContinual harness: online adaptation for self\-improving foundation agents\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2605.09998),[Link](https://arxiv.org/abs/2605.09998)Cited by:[§2\.3](https://arxiv.org/html/2610.07250#S2.SS3.p1.1)\.
- Kovalevet al\.\(2025\)V\. Kovalev, A\. Kuvshinov, A\. Buzovkin, D\. Pokidov, and D\. TimoninCRAFT: continuous reasoning and agentic feedback tuning for multimodal text\-to\-image generation\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2512.20362),[Link](https://arxiv.org/abs/2512.20362)Cited by:[§2\.1](https://arxiv.org/html/2610.07250#S2.SS1.p1.1)\.
- Liet al\.\(2016\)J\. Li, M\. Galley, C\. Brockett, J\. Gao, and B\. DolanA diversity\-promoting objective function for neural conversation models\.InProceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,pp\. 110–119\.External Links:[Link](http://dx.doi.org/10.18653/v1/N16-1014),[Document](https://dx.doi.org/10.18653/v1/n16-1014)Cited by:[§E\.3](https://arxiv.org/html/2610.07250#A5.SS3.SSS0.Px4.p1.1)\.
- Liuet al\.\(2025\)J\. Liu, G\. Liu, J\. Liang, Y\. Li, J\. Liu, X\. Wang, P\. Wan, D\. Zhang, and W\. OuyangFlow\-grpo: training flow matching models via online rl\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2505.05470),[Link](https://arxiv.org/abs/2505.05470)Cited by:[§2\.2](https://arxiv.org/html/2610.07250#S2.SS2.p1.1)\.
- Niuet al\.\(2025\)Y\. Niu, M\. Ning, M\. Zheng, W\. Jin, B\. Lin, P\. Jin, J\. Liao, C\. Feng, F\. Meng, K\. Ning, B\. Zhu, and L\. YuanWISE: a world knowledge\-informed semantic evaluation for text\-to\-image generation\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2503.07265),[Link](https://arxiv.org/abs/2503.07265)Cited by:[§E\.4](https://arxiv.org/html/2610.07250#A5.SS4.p1.1),[§4\.1](https://arxiv.org/html/2610.07250#S4.SS1.SSS0.Px1.p1.1)\.
- Ruizet al\.\(2023\)N\. Ruiz, Y\. Li, V\. Jampani, Y\. Pritch, M\. Rubinstein, and K\. AbermanDreamBooth: fine tuning text\-to\-image diffusion models for subject\-driven generation\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),pp\. 22500–22510\.Cited by:[§2\.2](https://arxiv.org/html/2610.07250#S2.SS2.p1.1)\.
- Smithet al\.\(2023\)J\. S\. Smith, Y\. Hsu, L\. Zhang, T\. Hua, Z\. Kira, Y\. Shen, and H\. JinContinual diffusion: continual customization of text\-to\-image diffusion with c\-lora\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2304.06027),[Link](https://arxiv.org/abs/2304.06027)Cited by:[§2\.2](https://arxiv.org/html/2610.07250#S2.SS2.p1.1)\.
- Wallaceet al\.\(2024\)B\. Wallace, M\. Dang, R\. Rafailov, L\. Zhou, A\. Lou, S\. Purushwalkam, S\. Ermon, C\. Xiong, S\. Joty, and N\. NaikDiffusion model alignment using direct preference optimization\.In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),pp\. 8228–8238\.External Links:[Link](http://dx.doi.org/10.1109/cvpr52733.2024.00786),[Document](https://dx.doi.org/10.1109/cvpr52733.2024.00786)Cited by:[§2\.2](https://arxiv.org/html/2610.07250#S2.SS2.p1.1),[§4\.1](https://arxiv.org/html/2610.07250#S4.SS1.SSS0.Px3.p1.1),[Table 2](https://arxiv.org/html/2610.07250#S4.T2.2.4.1.1.1)\.
- Wanget al\.\(2025\)K\. Wang, R\. Chen, T\. Zheng, and H\. HuangImAgent: a unified multimodal agent framework for test\-time scalable image generation\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2511.11483),[Link](https://arxiv.org/abs/2511.11483)Cited by:[§1](https://arxiv.org/html/2610.07250#S1.p1.1),[§2\.1](https://arxiv.org/html/2610.07250#S2.SS1.p1.1)\.
- Wuet al\.\(2026\)J\. Wu, S\. Yang, Z\. Lu, F\. Zhang, Y\. Shen, L\. Feng, H\. Luo, Z\. Lian, S\. Zhang, Z\. Wen, and J\. TaoSEED: self\-evolving on\-policy distillation for agentic reinforcement learning\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2607.14777),[Link](https://arxiv.org/abs/2607.14777)Cited by:[§2\.3](https://arxiv.org/html/2610.07250#S2.SS3.p1.1)\.
- Xiaet al\.\(2026a\)P\. Xia, J\. Chen, H\. Wang, J\. Liu, K\. Zeng, Y\. Wang, S\. Han, Y\. Zhou, X\. Zhao, H\. Chen, Z\. Zheng, C\. Xie, and H\. YaoSkillRL: evolving agents via recursive skill\-augmented reinforcement learning\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2602.08234),[Link](https://arxiv.org/abs/2602.08234)Cited by:[§2\.3](https://arxiv.org/html/2610.07250#S2.SS3.p1.1)\.
- Xiaet al\.\(2026b\)P\. Xia, J\. Chen, X\. Yang, H\. Tu, J\. Liu, K\. Xiong, S\. Han, S\. Qiu, H\. Ji, Y\. Zhou, Z\. Zheng, C\. Xie, and H\. YaoMetaClaw: just talk – an agent that meta\-learns and evolves in the wild\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2603.17187),[Link](https://arxiv.org/abs/2603.17187)Cited by:[§2\.3](https://arxiv.org/html/2610.07250#S2.SS3.p1.1)\.
- Xuet al\.\(2023\)J\. Xu, X\. Liu, Y\. Wu, Y\. Tong, Q\. Li, M\. Ding, J\. Tang, and Y\. DongImageReward: learning and evaluating human preferences for text\-to\-image generation\.InAdvances in Neural Information Processing Systems 36,NeurIPS 2023,pp\. 15903–15935\.External Links:[Link](http://dx.doi.org/10.52202/075280-0700),[Document](https://dx.doi.org/10.52202/075280-0700)Cited by:[§2\.2](https://arxiv.org/html/2610.07250#S2.SS2.p1.1)\.
- Yeet al\.\(2026\)T\. Ye, L\. Dong, X\. Wu, S\. Huang, and F\. WeiOn\-policy context distillation for language models\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2602.12275),[Link](https://arxiv.org/abs/2602.12275)Cited by:[§1](https://arxiv.org/html/2610.07250#S1.p4.1),[§3\.2\.1](https://arxiv.org/html/2610.07250#S3.SS2.SSS1.p2.1)\.
- Zhanget al\.\(2026a\)H\. Zhang, Q\. Long, J\. Bao, T\. Feng, W\. Zhang, H\. Yue, and W\. WangMemSkill: learning and evolving memory skills for self\-evolving agents\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2602.02474),[Link](https://arxiv.org/abs/2602.02474)Cited by:[§2\.3](https://arxiv.org/html/2610.07250#S2.SS3.p1.1)\.
- Zhanget al\.\(2026b\)Z\. Zhang, J\. Li, J\. Zhang, K\. Gao, K\. Yan, L\. Jiang, N\. Tang, S\. Yin, T\. Wu, X\. Chen, X\. Xu, Y\. Shu, Y\. Zhang, Y\. Xu, Y\. Chen, Z\. Wang, Z\. Liu, Z\. Zhou, H\. Zhang, D\. Zhao, and C\. WuQwen\-image\-agent: bridging the context gap in real\-world image generation\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2606.26907),[Link](https://arxiv.org/abs/2606.26907)Cited by:[§2\.1](https://arxiv.org/html/2610.07250#S2.SS1.p1.1)\.
- Zhanget al\.\(2026c\)Z\. Zhang, Y\. Lin, N\. L\. Kuang, L\. Wu, X\. Li, S\. Liu, and F\. MaCo\-evolving skill generation and policy optimization\.arXiv\.External Links:[Document](https://dx.doi.org/10.48550/ARXIV.2606.08755),[Link](https://arxiv.org/abs/2606.08755)Cited by:[§2\.3](https://arxiv.org/html/2610.07250#S2.SS3.p1.1)\.
- Zhaoet al\.\(2024\)A\. Zhao, D\. Huang, Q\. Xu, M\. Lin, Y\. Liu, and G\. HuangExpeL: llm agents are experiential learners\.Proceedings of the AAAI Conference on Artificial Intelligence38\(17\),pp\. 19632–19642\.External Links:ISSN 2159\-5399,[Link](http://dx.doi.org/10.1609/aaai.v38i17.29936),[Document](https://dx.doi.org/10.1609/aaai.v38i17.29936)Cited by:[§2\.3](https://arxiv.org/html/2610.07250#S2.SS3.p1.1)\.

## Appendix AQualitative Examples for the Main Results

Figure[3](https://arxiv.org/html/2610.07250#A1.F3)shows selected test queries for which D\-OPCD improves direct generation\. Table[1](https://arxiv.org/html/2610.07250#S4.T1)reports the aggregate results over the full evaluation sets\.

![Refer to caption](https://arxiv.org/html/2610.07250v1/main_results_qualitative.png)Figure 3:Matched outputs for four test queries\. The left triplet uses the base generatorπθ\\pi\_\{\\theta\}and the right triplet uses the D\-OPCD\-updated generatorπθ′\\pi\_\{\\theta^\{\\prime\}\}\. Within each triplet, the columns show direct generation, the skill\-free harnessℋ∅\\mathcal\{H\}\_\{\\emptyset\}, and evolved skills \(ℋr⋆\\mathcal\{H\}\_\{r^\{\\star\}\}orℋr†\\mathcal\{H\}\_\{r^\{\\dagger\}\}\)\. The text below each row is the original query; harness runs may use selected generation prompts\.
## Appendix BDiscussion on Limitations and Future Works

##### Limits of internalizing skill\-evolution experience\.

Internalizing the experience from skill evolution yields only modest gains for the standalone generator: D\-OPCD raises the four\-benchmark direct\-generation average from 60\.52 to 65\.09, still far below the 81\.83 achieved by the original skill\-equipped harness \(Table[1](https://arxiv.org/html/2610.07250#S4.T1)\)\. One possible explanation is that prompts produced by evolved skills, while effective when used by the harness, may be poorly aligned with the diffusion model’s latent space for learning the same behavior from the original queryqiq\_\{i\}alone\. Further studies could explore which harness designs produce experience that supports both immediate task success and subsequent model learning, and which algorithms can internalize this broader harness experience\. Stronger mechanisms for jointly adapting the harness and generator may be needed to realize sustained co\-evolution\.

##### Evidence for repeated co\-evolution\.

Our experiments evaluate one base generator and one D\-OPCD update on four benchmarks, followed by one subsequent round of skill evolution\. They show that a newly evolved harness can build on the updated generator, but do not establish what happens over many update\-and\-reset cycles or across different generators and task distributions\. Longer\-term studies could measure retention of earlier gains, transfer to new tasks, and the tradeoff between offline training cost and recurring harness inference cost\. Repeated model updates may also require methods that prevent forgetting while adapting the harness to each new generator\.

##### Beyond text\-to\-image agents\.

The agent studied here improves generation through prompt construction, verification, and refinement, without external knowledge search or reference\-image tools\. Other image\-generation agents may search the web, retrieve reference images, and use additional tools\. In addition to text\-to\-image generation, they may handle image\-to\-image generation and other image\-generation tasks\. Their experience includes tool choices, retrieved information, and visual evidence that a final text prompt cannot fully represent\. Extending internalization to these settings requires learning which reusable capabilities can be transferred into model weights and which information should remain available through tools at inference\. A further open problem is how to use tool\-augmented experience gathered at test time to guide model updates and then evolve the harness again around the updated generator\.

## Appendix CImplementation Details of the Agent Harness and Skill Evolution

### C\.1GEMS\-Based Agent Harness Design

##### Planning and requirements\.

We adapt the GEMS agent architecture\([He et al\., 2026](https://arxiv.org/html/2610.07250#bib.bib3)\)to a text\-to\-image harness with five roles: Planner, Decomposer, Generator, Verifier, and Refiner\. Given an original queryqiq\_\{i\}, the Planner selects applicable Skills using their names and descriptions, then loads their instructions to construct the first generation promptpi,1p\_\{i,1\}\. If it selects no Skill,pi,1=qip\_\{i,1\}=q\_\{i\}\. The Decomposer then turnsqiq\_\{i\}into a fixed set of yes/no visual requirements𝒞i=\(ci,1,…,ci,mi\)\\mathcal\{C\}\_\{i\}=\(c\_\{i,1\},\\ldots,c\_\{i,m\_\{i\}\}\)\. Deriving these checks from the original query keeps the target of verification fixed as prompts are revised\.

##### Generation and refinement\.

The Generator produces an imageIi,jI\_\{i,j\}from the current promptpi,jp\_\{i,j\}\. The multimodal Verifier checks the image against every requirement in𝒞i\\mathcal\{C\}\_\{i\}\. If any check fails, the Refiner uses the prompt, image, verification feedback, and task\-local history of earlier attempts to producepi,j\+1p\_\{i,j\+1\}\. That history retains the prompts, images, check results, and short experience summaries used during refinement\. The agent stops when all checks pass or the attempt budget is exhausted\. During evolution, relevant Insights may also be retrieved to guide refinement after a failure; frozen skill\-only replay and evaluation disable this retrieval\. External benchmark scores do not guide refinement within a task\.

##### Returned output\.

Letni,jn\_\{i,j\}be the number of requirements satisfied by attemptjj, and letJiJ\_\{i\}be the number of attempts executed\. The agent returns the prompt and image from the earliest attempt with the highest count:

ji∗=min⁡\{j:ni,j=max1≤ℓ≤Ji⁡ni,ℓ\},\(pi,Ii\)=\(pi,ji∗,Ii,ji∗\)\.j\_\{i\}^\{\*\}=\\min\\left\\\{j:n\_\{i,j\}=\\max\_\{1\\leq\\ell\\leq J\_\{i\}\}n\_\{i,\\ell\}\\right\\\},\\qquad\(p\_\{i\},I\_\{i\}\)=\(p\_\{i,j\_\{i\}^\{\*\}\},I\_\{i,j\_\{i\}^\{\*\}\}\)\.\(10\)Thus, a later unsuccessful revision can leave an earlier image as the returned output\. The resulting\(pi,Ii\)\(p\_\{i\},I\_\{i\}\)is the harness output used for context\-data collection\. The agent prompt templates are listed in Appendix[C\.5](https://arxiv.org/html/2610.07250#A3.SS5)\.

### C\.2Episode Construction and Insight Extraction

##### Episode records\.

Each completed task yields one Episode\. An Episode records \(i\) the task metadata, including the original requestqiq\_\{i\}and the deployed generator version; \(ii\) the ordered generation attempts and verification results; \(iii\) the Skills selected at the start of the task and any Insights retrieved for refinement; and \(iv\) the returned attempt, its passed and failed checks, the first\-attempt check score, and any feedback from the external evaluator\. Multiple attempts within the same Episode still count as one evidence unit\. An optional LLM\-written summary may be attached, but we disable it so that consolidation uses the factual attempt record and outcome\.

##### Frozen working set\.

At the batch barrier, the learner builds one retrieval query per Episode from its request and failure evidence, retrieves the top\-44active Insights whose cosine similarity to the query exceeds0\.350\.35using BGE\-M3 embeddings, and takes the union over Episodes as the working set𝒩k\\mathcal\{N\}\_\{k\}\. This management\-only retrieval includes immature single\-Episode hypotheses, so that new evidence can strengthen, correct, or refute them\. The working set is frozen before the manager runs: the Insight managerℳI\\mathcal\{M\}\_\{I\}may reference only these entries and only Episodes of the current batch\.

##### Action space\.

ℳI\\mathcal\{M\}\_\{I\}returns an ordered list of operations\. Table[6](https://arxiv.org/html/2610.07250#A3.T6)summarizes the six actions, and Figure[10](https://arxiv.org/html/2610.07250#A3.F10)shows the manager’s decision prompt\. The action space is deliberately small: evidence moves are separated from text moves, so support and contradiction accumulate as Episode references while wording changes remain rare, explicit, and evidence\-backed\.

Table 6:Action space of the Insight managerℳI\\mathcal\{M\}\_\{I\}\. All actions require at least one Episode reference from the current batch; targets must be active Insights in the frozen working set\.
##### Validation and commit\.

Operations are checked in order against the frozen working set: references must resolve, text constraints must hold per action, and one Episode may not take both a supporting and a contradicting direction toward the same Insight\. Only operations satisfying these rules update the Insight pool\.

##### Evidence and maturity\.

Support and contradiction counts use distinct Episodes, not the number of attempts within an Episode\. The support thresholdτ\\tauand the minimum number of supporting batches required before an Insight is passed to the Skill manager vary by benchmark \(Table[9](https://arxiv.org/html/2610.07250#A3.T9)\)\. The Insight store retains contradictory Episode references so they can affect maturity and motivate a revision or archival\.

### C\.3Skill Updates and Library Versioning

##### Update trigger\.

After each batchkk, Skill evolution is considered when at leastBBmature, previously unreviewed Insights have accumulated and the minimum interval since the preceding completed Skill review has elapsed\. Both settings vary by benchmark \(Table[9](https://arxiv.org/html/2610.07250#A3.T9)\)\. The mature\-insight trigger controls Skill maintenance independently of generator updates\.

##### Action space\.

Once triggered, the Skill managerℳS\\mathcal\{M\}\_\{S\}examines the pending mature Insights together with summaries of the current library and the recent change history, and returns one ordered plan from the nine\-action space in Table[7](https://arxiv.org/html/2610.07250#A3.T7)\. Figure[11](https://arxiv.org/html/2610.07250#A3.F11)shows its decision prompt\. Text\-level edits \(Replace,Insert,Delete\) operate on exact, uniquely matching passages copied from the target Skill, which makes every proposed diff verifiable before it is applied\. Structural actions \(Split,Merge,Retire\) keep the library routeable under a hard capacity bound\. Every edit cites the Insights it absorbs, linking each Skill revision back to its evidence\.

Table 7:Action space of the Skill managerℳS\\mathcal\{M\}\_\{S\}\.source\_insight\_refsidentifies the mature Insights absorbed by an edit\. An empty plan is a deliberate no\-op\.
##### Validation and commit\.

Proposed plans are checked against two hard constraints: at mostC=5C\{=\}5active Skills, and at mostL=4500L\{=\}4500characters per Skill\. Only a valid final library is committed\. Each committed edit records the Insights it absorbs\.

##### Reviewed evidence and harness versions\.

The review state is recorded per Insight at its latest evidence revision\. An accepted Skill update marks the Insights it absorbs as reviewed, preventing the same unchanged evidence from immediately triggering another update\. A deliberateNoOpalso completes a review but changes neither the active library nor its version\. If a reviewed Insight later gains new evidence, its new revision can be considered in a subsequent update\. Each committed library change defines a new harness version\.

##### Deployment\-time use\.

At inference the agent routes each new request to the relevant Skills once, before the first generation\. Generation calls under the evolving library are recorded with their harness and generator version annotations and become the prompt–image pairs available to D\-OPCD\.

### C\.4Experimental Configuration and Frozen Replay

##### Experimental configuration\.

GPT\-5\.6 Terra with medium reasoning effort serves as the agent’s base reasoning model, the Insight and Skill managers, and the in\-task verifier\. Table[8](https://arxiv.org/html/2610.07250#A3.T8)reports settings shared across benchmarks; Table[9](https://arxiv.org/html/2610.07250#A3.T9)reports benchmark\-specific ASE settings\. Evaluation leaves the agent state unchanged, so it cannot affect subsequent predictions or the adaptation stream\.

Table 8:Harness\-evolution settings shared across benchmarks\.ComponentParameterValueAgent loopmax verification questions1010Agent loopmax generation attempts55Episode storeLLM summarydisabledInsight extractiontask batch sizeMM88Insight retrievaltext embedderBGE\-M3Insight retrievaltop\-kk/ semantic gate44/0\.350\.35cosineInsight consolidationduplicate threshold0\.920\.92cosineSkill librarymax active SkillsCC55Skill librarymax Skill lengthLL45004500charactersTable 9:Benchmark\-specific ASE settings during evolution on𝒬train\\mathcal\{Q\}\_\{\\mathrm\{train\}\}\. The review interval is measured in task batches\.ParameterGenEvalGenEval2WISER2I\-BenchInsight supportτ\\tau22444444Min\. support batches22333333Skill triggerBB33888888Min\. review interval44666666
##### Generation seeds\.

For taskii, generation attemptjjuses seedsi\+j−1s\_\{i\}\+j\-1, wheresis\_\{i\}is the task’s base seed\.

##### Selecting and freezing a skill library\.

Library selection is performed on𝒬hold\\mathcal\{Q\}\_\{\\mathrm\{hold\}\}, independently of evaluation on𝒬eval\\mathcal\{Q\}\_\{\\mathrm\{eval\}\}\. Each candidate is evaluated without persistent updates; the selected library is subsequently frozen for replay on the entire𝒬train\\mathcal\{Q\}\_\{\\mathrm\{train\}\}and for evaluation on𝒬eval\\mathcal\{Q\}\_\{\\mathrm\{eval\}\}\. The replayed tasks do not update the Skill library or the persistent Insight store\. Table[10](https://arxiv.org/html/2610.07250#A3.T10)identifies which Skill\-update snapshot was selected in each evolution round and how many tasks had contributed experience to that snapshot\. The skill\-free harnessℋ∅\\mathcal\{H\}\_\{\\emptyset\}uses the empty library and requires no version selection\.

Table 10:Selected Skill\-update snapshots in both evolution rounds\. The selected update isr⋆r^\{\\star\}in Round 1 andr†r^\{\\dagger\}in Round 2;nrn\_\{r\}counts completed tasks contributing experience to that snapshot\. Selection scores use GenEval2 Soft\-TIFA AM and the other benchmarks’ native metrics\.
##### Prompt\-pair collection\.

For each queryqi∈𝒬trainq\_\{i\}\\in\\mathcal\{Q\}\_\{\\mathrm\{train\}\}, frozen replay returns\(pi,Ii\)\(p\_\{i\},I\_\{i\}\)according to Equation[10](https://arxiv.org/html/2610.07250#A3.E10)\. We retain the submitted record\(qi,pi,Ii\)\(q\_\{i\},p\_\{i\},I\_\{i\}\); all compared training methods drop tasks whose prompt equals the query after normalization, following the context\-data protocol in Appendix[D\.1](https://arxiv.org/html/2610.07250#A4.SS1)\.

### C\.5Prompt Templates

#### C\.5\.1Agent Execution Prompts

Figures[4](https://arxiv.org/html/2610.07250#A3.F4)–[9](https://arxiv.org/html/2610.07250#A3.F9)show the prompts used by the agent during each generation task in execution order\. Braced capitalized fields denote runtime substitutions, and line wrapping is for presentation only\. The optional cross\-task Insight field in the refinement template is used during evolution and left empty in frozen skill\-only replay and evaluation\.

Youareanimage\-generationSkillRouter\.EveryavailableSkillisaninitial\-promptrewriterthatmayrunexactlyoncebeforethefirstimagegeneration\.ChoosethesubsetofcandidateSkillswhosetransformationsareapplicableandusefulfortheoriginalrequest\.YoumaycombinemultiplecomplementarySkillsorchoosenone\.

\#\#\#CandidateInitial\-PromptSkills

\{SKILL\_MANIFEST\}

\#\#\#OriginalUserRequest

\{ORIGINAL\_QUERY\}

ReturnONLYaJSONarraycontainingtheselectedSKILL\_IDstringsintheordertheyshouldbeconsidered\.Return\[\]whennoSkillisuseful\.

Figure 4:Skill routing prompt\.Selects applicable Skills from their manifest before the first generation attempt\.Producetheonepromptusedforthefirstimage\-generationattempt\.Preserveeveryrequestedfact\.Applyallselectedinitial\-promptSkillinstructionstogetherasonecoherenttransformation\.Combinecompatibleguidanceaccordingtotheoriginalrequestandreturnonlytheresultingprompt\.

\#\#\#Initial\-PromptSkillInstructions

\{SELECTED\_SKILL\_INSTRUCTIONS\}

\#\#\#OriginalPrompt

\{ORIGINAL\_QUERY\}

ReturnONLYthefinalenhancedprompt\.

Figure 5:Initial prompt rewrite\.Combines selected Skill instructions with the original query\.Analyzetheuser’simagegenerationpromptandbreakitintospecificvisualrequirements\.Foreachrequirement,writeaquestionanswerablewithyesorno\.Thequestionsmustverifywhethertherequirementispresentinanimage\.Produceatmost10independentquestions\.Prioritizethemostimportantrequirementsandconstraintsthatanimagegeneratorislikelytomiss\.

RespondONLYwithaJSONarrayofstrings\.

USERPROMPT:

\{ORIGINAL\_QUERY\}

Figure 6:Requirement decomposition\.Converts the original query into visual yes/no checks\.Image:<image\>

Evaluateeveryverificationquestionbelowagainsttheprovidedimage\.ReturnONLYoneJSONarraywithexactlythesamenumberofitemsandinthesameorderasthequestions\.Eachitemmustbeanobjectwithan’answer’stringanda’passed’JSONboolean\.Usetrueonlywhentheimageclearlysatisfiesthequestion;otherwiseusefalse\.DonotusemarkdownoraddcommentaryoutsidetheJSONarray\.

VERIFICATIONQUESTIONS:

1\.\{QUESTION\_1\}

2\.\{QUESTION\_2\}

…

Figure 7:Batched image verification\.Checks an image against all visual requirements in one request\.Task:Summarizetheexperienceofthecurrentimagegenerationattempt\.

—CURRENTATTEMPT—

Promptused:\{CURRENT\_PROMPT\}

Passedrequirements:\{PASSED\_REQUIREMENTS\}

Failedrequirements:\{FAILED\_REQUIREMENTS\}

Reasoning/Thoughtbeforegeneration:\{CURRENT\_THOUGHT\}

Image:<image\>

—PREVIOUSEXPERIENCES—

\{PREVIOUS\_TASK\_LOCAL\_EXPERIENCES\}

—ANALYSIS—

Basedontheimage,internalverificationresults,thoughtprocess,andpriorattemptsfromthistask,summarizewhatworked,whatfailed,andwhatshouldbechangedinthenextattempt\.Keepitunder100words\.Donotincludeanintroduction\.

Figure 8:Task\-local experience summary\.Records lessons from the current attempt and earlier attempts on the same task\.Task:Refinetheimagegenerationpromptafterfailedinternalchecks\.

ORIGINALINTENT:

\{ORIGINAL\_QUERY\}

—RELEVANTCROSS\-TASKINSIGHTS—

\{RELEVANT\_INSIGHTS\_IF\_ENABLED\}

—ATTEMPTHISTORY—

\{ATTEMPT\_HISTORY\}

—REQUIREMENTS—

1\.Reinforcerequirementsthatfailedinthelatestattempt\.

2\.Preserverequirementsthatpreviouslypassed\.

3\.Applyarelevantinsightonlywhenitsconditionmatches\.

4\.Keepalloriginaluserconstraintsandavoidconflictinglanguage\.

ReturnONLYthenewprompt\.

Figure 9:Feedback\-based refinement\.Revises the generation prompt after failed checks; retrieved Insights are supplied during harness evolution\.
#### C\.5\.2Harness Evolution Prompts

The Insight and Skill managers use separate prompts to update cross\-task knowledge\. Figures[10](https://arxiv.org/html/2610.07250#A3.F10)and[11](https://arxiv.org/html/2610.07250#A3.F11)show their core decision instructions and runtime inputs\. The operation schemas are given in Tables[6](https://arxiv.org/html/2610.07250#A3.T6)and[7](https://arxiv.org/html/2610.07250#A3.T7), rather than repeated in the prompt figures\.

MaintainthesituationalInsightmemoryusingthisbatchofEpisodesandtherelatedactiveInsightsretrievedforit\.

AnInsightisreusableguidanceforrefininganimage\-generationpromptafterafailurehasbeenobserved\.Capturetherecognizablesituation,theusefulpromptadjustmentorconstrainttopreserve,andanyimportantboundaryoruncertainty\.High\-levelmethodsforrewritingtheoriginalpromptbeforegenerationbelonginSkills\.

Buildthesmallestcoherentmemoryupdatesupportedbytheevidence\.Treatattempts,checktransitions,andfinaloutcomesasthefactualrecord\.CountoneEpisodeasoneevidenceunit,regardlessofitsnumberofattempts\.

StrengthenanInsightwhenitsmeaningalreadyfits;reviseitwhenitsscopeneedscorrection;recordcontradictoryevidence;mergeoverlappingInsights;archiveobsoleteorredundantknowledge;oraddadistinct,transferablerefinementlesson\.Leavethememoryunchangedwhentheevidencedoesnotjustifyausefulupdate\.

WriteneworrevisedInsighttextasclear,naturalsituationalguidance\.ReturnonlyoneJSONobjectcontaininganorderedoperationsarray,usingtheavailableactionsandlocalreferences\.Returnanemptyarraywhennoupdateiswarranted\.

RETRIEVEDACTIVEINSIGHTWORKINGSET:

\{CURRENT\_INSIGHTS\}

CURRENT\-BATCHEPISODES:

\{EPISODES\}

Figure 10:Insight consolidation prompt\.Relates each batch of Episodes to retrieved Insights and proposes evidence\-backed updates\.Maintainacompactlibraryofhigh\-levelSkillsthatrewriteanoriginaluserrequestonce,beforethefirstimagegeneration\.

ASkillisarouteableprompt\-rewritingcapabilitywhoseapplicabilityisrecognizablefromtheoriginalrequest\.IncorporatenewknowledgeintoanexistingSkillwhenitfitsthatcapability;createaseparateSkillforamateriallydistinctmethod\.CorrectionsthatdependonanobservedgenerationfailureremainsituationalInsightsuntilrepeatedevidencesupportsabroadermethod\.

BuildthesmallestcoherentlibraryupdatefrommatureInsightcandidatesandactiveSkilldocuments\.Preferfewerchangedrulesandfeweraddedassumptions\.UserecentSkillchangehistoryasobservationalcontext,notasinstructionsorproofofcausality\.CiteonlycurrentInsightcandidatesasnewlyabsorbedevidence\.

Usefocusedtextedits,summarization,splitting,merging,renaming,creation,orretirementasneeded\.Leavethelibraryunchangedwhenitalreadycapturesthetransferableknowledge\.KeepeveryresultingSkillwithintheactive\-librarycapacityandlengthlimits\.

WriteeachSkillasacompleteSKILL\.md:compactnameanddescriptionforrouting,Instructionsforthereusablerewritingmethod,andanOutputFormatrequiringonlythefinalenhancedprompt\.ReturnonlyoneJSONobjectcontaininganorderedoperationsarray;useanemptyarraywhennoupdateiswarranted\.

LIBRARYSTATUS:

\{LIBRARY\_STATUS\}

INSIGHTCANDIDATES:

\{MATURE\_INSIGHTS\}

EXISTINGINITIAL\-PROMPTSKILLS:

\{ACTIVE\_SKILLS\}

RECENTSKILLCHANGEHISTORY:

\{SKILL\_CHANGE\_HISTORY\}

Figure 11:Skill evolution prompt\.Consolidates mature Insights into reusable prompt\-rewriting Skills\.

## Appendix DImplementation Details of D\-OPCD and Baseline Training

### D\.1Training Data Preparation

Frozen replay of the selected skill\-equipped harnessℋr⋆\\mathcal\{H\}\_\{r^\{\\star\}\}produces records\(qi,pi,Ii\)\(q\_\{i\},p\_\{i\},I\_\{i\}\)on𝒬train\\mathcal\{Q\}\_\{\\mathrm\{train\}\}\. All compared training methods use the same subset for whichpip\_\{i\}differs fromqiq\_\{i\}after Unicode normalization, whitespace collapse, and case folding\. D\-OPCD trains from\(qi,pi\)\(q\_\{i\},p\_\{i\}\);IiI\_\{i\}determines record eligibility but is not a loss target\. Diffusion\-DPO additionally pairs the submitted imageIiI\_\{i\}with a rejected image generated directly fromqiq\_\{i\}using the same seed\. Vanilla SFT and D\-OPSD train from the corresponding\(qi,Ii\)\(q\_\{i\},I\_\{i\}\)records\. The per\-benchmark counts and filtering yields are reported in Appendix[E\.2](https://arxiv.org/html/2610.07250#A5.SS2)\.

### D\.2D\-OPCD Implementation

The student receivesqiq\_\{i\}\. For the teacher, the text encoder receivesqiq\_\{i\}followed by a blank line, the labelPrivileged generation context:, andpip\_\{i\}\. The student and teacher use the same frozen text encoder, with maximum sequence lengths of 512 and 1024 tokens, respectively\. Starting from Gaussian noise, the student follows a four\-point Euler rollout att∈\{0,0\.1,0\.25,0\.5\}t\\in\\\{0,0\.1,0\.25,0\.5\\\}, with its final update ending att=1t=1\. The state is detached before each velocity comparison, so gradients update the student at visited states without backpropagating through earlier rollout steps\. An EMA copy of the student, with decay0\.99990\.9999, supplies the teacher prediction under the additional context\. We minimize the mean squared velocity difference across the four states, as defined in Section[3\.2](https://arxiv.org/html/2610.07250#S3.SS2); the teacher prediction is stop\-gradient\.

### D\.3Baseline Objectives

##### Vanilla SFT\.

The student receivesqiq\_\{i\}, while the agent\-submitted imageIiI\_\{i\}supplies the clean image latent for the standard flow\-matching target\. Each batch samples a noise level from a logit\-normal distribution with logit mean zero and standard deviation one\. No privileged prompt enters the model\.

##### Diffusion\-DPO\.

Both the chosen imageIiI\_\{i\}and the same\-seed rejected image are trained underqiq\_\{i\}\. A frozen copy of the initial generator supplies reference losses, and the chosen–rejected difference in model and reference flow\-matching losses forms the preference objective withβ=2500\\beta=2500\. Two logit\-normal noise levels are sampled per pair\.

##### D\-OPSD\.

The student receivesqiq\_\{i\}and visits states along its own sampling trajectory\. The EMA teacher, updated with decay0\.99990\.9999, receives an embedding of\(qi,Ii\)\(q\_\{i\},I\_\{i\}\)from Qwen3\-VL\-4B\-Instruct and supplies a detached clean\-output endpoint target\. This baseline shares SFT’s image\-paired records, but uses on\-policy student states rather than the image\-derived states used by SFT\.

### D\.4Optimization and Checkpoint Selection

All methods start from Z\-Image\-Turbo and train LoRA adapters on its transformer feed\-forward projections and attention query, key, value, and output projections\. The common settings are rank 64, scale parameterα=128\\alpha=128, zero LoRA dropout, 1024\-pixel square images, and seed 42\. We use AdamW withβ1=0\.9\\beta\_\{1\}=0\.9,β2=0\.999\\beta\_\{2\}=0\.999, zero weight decay, a constant learning rate of10−410^\{\-4\}without warmup, gradient clipping at norm 1, and bfloat16 precision\. Training runs for 2000 optimizer steps and saves a checkpoint every 200 steps\. Direct evaluation uses eight sampling steps and guidance scale zero\.

##### Compute resources\.

Model training and image\-generation inference used NVIDIA A100 GPUs \(80 GB each\)\. Training jobs allocated four GPUs; inference used one GPU per generation worker, with multiple workers run in parallel where applicable\.

Each saved checkpoint is evaluated on𝒬hold\\mathcal\{Q\}\_\{\\mathrm\{hold\}\}; the benchmark\-specific selection metrics are defined in Appendix[E](https://arxiv.org/html/2610.07250#A5)\. Table[11](https://arxiv.org/html/2610.07250#A4.T11)lists each benchmark and compared method using skill\-equipped harness data, with the reported optimizer step and held\-out score\. Hold scores use the same benchmark\-native metric used for checkpoint evaluation and for the main evaluation tables\.

Table 11:Generator checkpoints and held\-out scores by method and benchmark\. All methods use records fromℋr⋆\\mathcal\{H\}\_\{r^\{\\star\}\}; scores use GenEval Overall, GenEval2 Soft\-TIFA AM, WIScore, and R2I\-Score \(0–100\)\.

## Appendix EExperimental and Evaluation Details

### E\.1Task Set Allocations

Table[12](https://arxiv.org/html/2610.07250#A5.T12)gives the disjoint task sets used for harness evolution, selection, and final evaluation\. GenEval, GenEval2, and WISE use approximately 50/15/35 allocations; R2I\-Bench uses a 780\-task experimental subset with its supplied split\.

Table 12:In\-distribution task allocations\. Ratios give the train/hold/evaluation shares of the task pool actually used\.BenchmarkPool\|𝒬train\|\|\\mathcal\{Q\}\_\{\\mathrm\{train\}\}\|\|𝒬hold\|\|\\mathcal\{Q\}\_\{\\mathrm\{hold\}\}\|\|𝒬eval\|\|\\mathcal\{Q\}\_\{\\mathrm\{eval\}\}\|Ratio \(%\)GenEval5532768319449\.9/15\.0/35\.1GenEval280040012028050/15/35WISE Verified100050015035050/15/35R2I\-Bench subset78048010020061\.5/12\.8/25\.6The first three splits use seed 42; GenEval2 and WISE preserve earlier task assignments, while R2I\-Bench uses its supplied split\.

### E\.2Replay Data Construction and Counts

Frozen replay on𝒬train\\mathcal\{Q\}\_\{\\mathrm\{train\}\}yields one submitted\(qi,pi,Ii\)\(q\_\{i\},p\_\{i\},I\_\{i\}\)record per task\. Table[13](https://arxiv.org/html/2610.07250#A5.T13)reports the records retained before and after excluding normalizedpi=qip\_\{i\}=q\_\{i\}pairs\. D\-OPCD, Diffusion\-DPO, Vanilla SFT, and D\-OPSD use the prompt\-changed tasks, with method\-specific supervision\. The replay uses the selected skill\-equipped harness version in Table[10](https://arxiv.org/html/2610.07250#A3.T10)\.

Table 13:Replay\-record yields from the selected skill\-equipped harnessℋr⋆\\mathcal\{H\}\_\{r^\{\\star\}\}\. The last column gives the prompt\-changed tasks used by all four compared training methods\.
### E\.3Context Effectiveness and Diversity Diagnostics

##### Improvement\-filtered pairs\.

The default set retains records from the selected skill\-equipped harness when the submitted prompt differs from the original query after normalization\. The improvement filter instead starts from all saved records, including unchanged prompts\. For each task, we pair the saved image from the harness’s selected attempt with the first image generated directly fromqiq\_\{i\}\. We take the requirement checklist𝒞i\\mathcal\{C\}\_\{i\}from the selected harness attempt and score both saved images against that same checklist using the harness’s internal verifier\. Previously recorded answers are reused when the checklist matches; otherwise, the direct\-query image is verified against𝒞i\\mathcal\{C\}\_\{i\}\. IfVi​\(I\)V\_\{i\}\(I\)is the number of passed checks for imageII, we retain\(qi,pi\)\(q\_\{i\},p\_\{i\}\)iffVi​\(Iip\)\>Vi​\(Iiq\)V\_\{i\}\(I\_\{i\}^\{p\}\)\>V\_\{i\}\(I\_\{i\}^\{q\}\)\. Missing or invalid verifier responses are excluded\. This comparison requires neither new image generation nor external benchmark scores\. Since the saved images may use different seeds, the filter selects observed verifier improvements rather than estimating a seed\-controlled prompt effect\. It selects 43, 264, 319, and 314 pairs on GenEval, GenEval2, WISE, and R2I\-Bench, respectively\. Three R2I\-Bench pairs havepi=qip\_\{i\}=q\_\{i\}but satisfy the image\-level criterion\.

##### Diversity\-adapted sampling\.

When the Skill router selects applicable Skills, the default harness combines their guidance withqiq\_\{i\}in one initial prompt rewrite\. To broaden the contexts produced by the same Skills, the diversity adapter acts at this step, before the first image\-generation attempt\. It samples one of four expression strategies at random for each task with selected Skills\.*Local edit*preserves the query’s wording and order where possible while adding relevant guidance\.*Restate*rewrites the request in new words\.*Integrate*weaves guidance into a single coherent image description\.*Separate*states the original request first and places guidance in separate sentences\. All four instruct the rewriter to preserve the task requirements rather than create variety by changing the requested content\. Table[14](https://arxiv.org/html/2610.07250#A5.T14)gives the exact strategy instructions, and Figure[12](https://arxiv.org/html/2610.07250#A5.F12)shows the shared rewrite prompt that receivesqiq\_\{i\}, the selected Skill instructions, and the sampled strategy\.

If no Skill is selected, the first attempt usesqiq\_\{i\}directly\. The Skill library, router, internal verifier, and later refinement steps remain the same, with at most five generation attempts per task\. We record the prompt of the agent’s final selected image aspip\_\{i\}, which need not be its first or last attempted prompt\. After excluding normalizedpi=qip\_\{i\}=q\_\{i\}cases, the adapted replay yields 276, 400, 491, and 381 pairs on GenEval, GenEval2, WISE, and R2I\-Bench\. All three context sets use the same D\-OPCD objective and student/\[qi;pi\]\[q\_\{i\};p\_\{i\}\]teacher interface\.

Table 14:Exact expression\-strategy instructions inserted into the diversity\-adapter rewrite template\.Producethepromptforthefirstimage\-generationattempt\.UsethesuppliedSkillsasguidanceforfulfillingtheoriginalrequest\.Applyguidanceonlywhereitisrelevanttothatrequest,andpreservetherequest’smeaningandrequirements\.IfaSkillconflictswiththeoriginalrequest,therequesttakespriority\.Returnonlytheresultingprompt\.Returningtheoriginalrequestunchangedisallowed\.

\#\#\#Initial\-PromptSkillInstructions

\{SELECTED\_SKILL\_INSTRUCTIONS\}

\#\#\#OriginalPrompt

\{ORIGINAL\_QUERY\}

\#\#\#Expressionstrategy

\{STRATEGY\_INSTRUCTION\}

Thisstrategychangesexpression,nottherequestedcontent\.Itdoesnotauthorizedroppingapplicableguidanceoraddingnewrequirementsmerelytocreatevariety\.

ReturnONLYthefinalenhancedprompt\.

Figure 12:Diversity\-adapter rewrite template\.The strategy instruction is chosen from Table[14](https://arxiv.org/html/2610.07250#A5.T14); all other text is shared across modes\.
##### Paired task effectiveness\.

For each retained\(qi,pi\)\(q\_\{i\},p\_\{i\}\), we generate one image fromqiq\_\{i\}and one frompip\_\{i\}using the same initial generator, seed, sampling settings, and task\. Both images are scored against the originalqiq\_\{i\}and its benchmark annotations\. LetSb​\(q\)S\_\{b\}\(q\)andSb​\(p\)S\_\{b\}\(p\)denote the benchmark\-native aggregate scores for settingbbon exactly its retained tasks, expressed on a 0–100 scale\. Table[15](https://arxiv.org/html/2610.07250#A5.T15)reportsΔb=Sb​\(p\)−Sb​\(q\)\\Delta\_\{b\}=S\_\{b\}\(p\)\-S\_\{b\}\(q\)in percentage points,100​Δb/Sb​\(q\)100\\Delta\_\{b\}/S\_\{b\}\(q\)as a relative percentage, and the percentage of individual tasks whosepip\_\{i\}image scores strictly above its pairedqiq\_\{i\}image\. GenEval uses Overall and GenEval2 uses Soft\-TIFA AM\. The same\-seed external evaluation is distinct from the agent’s internal\-verifier criterion used for filtering\.

Table 15:Effectiveness of the three training\-pair settings on their retained train tasks\.S⁡\(q\)S\(q\)andS⁡\(p\)S\(p\)use the original query as the scoring target: GenEval Overall, GenEval2 Soft\-TIFA AM, WIScore, or R2I\-Score, respectively\.Δ=S⁡\(p\)−S⁡\(q\)\\Delta=S\(p\)\-S\(q\)is in percentage points; Rel\. is100​Δ/S⁡\(q\)100\\Delta/S\(q\), and Win is the percentage of individual pairs with a strictly higherpp\-image score\. The external benchmark scorer evaluates both images in each pair\.
##### Context\-text diversity\.

We jointly fit character 3–5\-gram TF–IDF features to all three settings within each benchmark\. We measure normalized character Vendi score\([Friedman and Dieng, 2023](https://arxiv.org/html/2610.07250#bib.bib31)\)\(Vendi divided by sample size\), mean within\-set pairwise cosine similarity, and corpus Distinct\-1\([Li et al\., 2016](https://arxiv.org/html/2610.07250#bib.bib32)\)\(unique words divided by all words\)\. Mean prompt length aids interpretation\. To reduce the effect of cohort size, Table[16](https://arxiv.org/html/2610.07250#A5.T16)uses equal\-size subsamples, fixed by the smallest setting in each benchmark: 43/264/319/314 for GenEval/GenEval2/WISE/R2I\-Bench\. For larger settings, it averages 100 subsamples drawn with seed 42\. These are text\-diversity proxies rather than image\-diversity scores\. The R2I\-Bench improvement\-filtered pool includes three unchangedpi=qip\_\{i\}=q\_\{i\}pairs; we retain them to match the model\-training data\.

A separate earlier matched\-400 GenEval2 audit compared the default SkillOnly contexts with those from the diversity adapter, fitting TF–IDF jointly to those two corpora\. It found normalized lexical Vendi scores of 0\.7402 and 0\.7921, respectively, and Distinct\-2 values of 0\.3533 and 0\.3862\. Because this two\-setting word\-level audit used a different feature space and sample protocol, its numbers are not inserted into Table[16](https://arxiv.org/html/2610.07250#A5.T16)\.

Table 16:Text diversity of the retained privileged contextspip\_\{i\}\. Normalized character Vendi and Distinct\-1 are higher for more varied contexts; mean pairwise character TF–IDF cosine is lower\. PoolNNis the number of retained pairs, and Evalnnis the equal\-size subsample\. Length is the mean number of words\.BenchmarkSettingPoolNNEvalnnChar Vendi↑\\uparrowChar cosine↓\\downarrowDistinct\-1↑\\uparrowLengthGenEvalDefault276430\.77540\.15190\.176896\.4Improvement filter43430\.78640\.14310\.1767116\.7Diversity adapter276430\.78410\.14590\.181783\.5GenEval2Default4002640\.49140\.13740\.0483189\.3Improvement filter2642640\.48160\.13860\.0476197\.4Diversity adapter4002640\.56640\.10940\.0590111\.0WISEDefault4903190\.75830\.07960\.1680112\.3Improvement filter3193190\.73390\.08560\.1592122\.9Diversity adapter4913190\.81330\.06100\.194084\.7R2I\-BenchDefault3923140\.76680\.07080\.1183143\.7Improvement filter3143140\.75820\.07210\.1135153\.3Diversity adapter3813140\.79460\.06320\.1265123\.5Character n\-grams probe repetition in the prompt wording that the adapter is designed to reduce\. A word unigram–bigram audit agrees on GenEval2, WISE, and R2I\-Bench, but its GenEval differences are small and mixed: normalized word Vendi is 0\.9221 for the default and 0\.9205 for the adapter, while Distinct\-2 is 0\.5469 and 0\.5430, respectively\.

### E\.4Evaluation Metrics and Protocols

For in\-distribution evaluation, each method generates one image per task on𝒬eval\\mathcal\{Q\}\_\{\\mathrm\{eval\}\}\. We report each benchmark by its native metric under its official evaluation protocol\. GenEval\([Ghosh et al\., 2023](https://arxiv.org/html/2610.07250#bib.bib12)\)uses the Overall score from its detector\-based checker; GenEval2\([Kamath et al\., 2025](https://arxiv.org/html/2610.07250#bib.bib20)\)uses prompt\-level Soft\-TIFA AM; WISE\([Niu et al\., 2025](https://arxiv.org/html/2610.07250#bib.bib21)\)uses WiScore under the WISE\_Verified binary protocol; and R2I\-Bench\([Chen et al\., 2025](https://arxiv.org/html/2610.07250#bib.bib22)\)uses R2I\-Score from its checklist\-based judge\. Checkpoint candidates are evaluated on𝒬hold\\mathcal\{Q\}\_\{\\mathrm\{hold\}\}, with Overall score for GenEval, Soft\-TIFA AM for GenEval2, WiScore for WISE, and R2I\-Score for R2I\-Bench\. Reported hold scores use these same evaluation metrics\.

T2I\-CompBench\+\+\([Huang et al\., 2025](https://arxiv.org/html/2610.07250#bib.bib23)\)is evaluated separately on 2,100 prompts from the seven categories in Table[5](https://arxiv.org/html/2610.07250#S4.T5)of its official*validation*split, using three images per prompt with seeds 42–44 and reporting scores by category\. This OOD set is not𝒬hold\\mathcal\{Q\}\_\{\\mathrm\{hold\}\}and is used neither for skill evolution nor for model or harness selection\. The base generator is evaluated once, and each other row in Table[5](https://arxiv.org/html/2610.07250#S4.T5)reports one source\-specific D\-OPCD checkpoint, without averaging across training sources\.

## Appendix FUse of AI Assistants

AI assistants were used in three roles\. First, the experimental harness uses GPT\-5\.6 Terra to plan and refine generation prompts, verify images, and consolidate task experience into Insights and Skills \(Appendix[C](https://arxiv.org/html/2610.07250#A3)\)\. Second, outputs are scored with the benchmark\-native automated evaluators described in Appendix[E](https://arxiv.org/html/2610.07250#A5), including the checklist\-based judge for R2I\-Bench\. Third, general\-purpose AI writing tools were used to improve wording and readability\. They were not used to originate the research questions, choose the experimental design, produce the reported results, or draw the paper’s conclusions\. The authors reviewed and approved the experimental design, analyses, results, and manuscript text\.